The Architecture of FunGramKB
Carlos Periñán-Pascual, Francisco Arcas-Túnez
Universidad Católica San Antonio
Campus de los Jerónimos s/n, 30107 Guadalupe (Murcia), Spain
E-mail: jcperinan@pdi.ucam.edu, farcas@pdi.ucam.edu
Abstract
Natural language understanding systems require a knowledge base provided with conceptual representations reflecting the structure of
human beings’ cognitive system. Although surface semantics can be sufficient in some other systems, the construction of a robust
knowledge base guarantees its use in most natural language processing applications, consolidating thus the concept of resource reuse.
In this scenario, FunGramKB is presented as a multipurpose knowledge base whose model has been particularly designed for natural
language understanding tasks. The theoretical basement of this knowledge engineering project lies in the construction of two
complementary types of interlingua: the conceptual logical structure, i.e. a lexically-driven interlingua which can predict linguistic
phenomena according to the Role and Reference Grammar syntax-semantics interface, and the COREL scheme, i.e. a concept-oriented
interlingua on which our rule-based reasoning engine is able to make inferences effectively. The objective of the paper is to describe
the different conceptual, lexical and grammatical modules which make up the architecture of FunGramKB, together with an
exploratory outline on how to exploit such a knowledge base within an NLP system.
1.
Introduction
FunGramKB Suite1 is a user-friendly online environment
for the semiautomatic construction of a multipurpose
lexico-conceptual knowledge base for natural language
processing (NLP) systems, and more particularly for
natural language understanding. On the one hand,
FunGramKB is multipurpose in the sense that it is both
multifunctional and multilingual. Thus, FunGramKB has
been designed to be potentially reused in many NLP tasks
(e.g. information retrieval and extraction, machine
translation, dialogue-based systems, etc) and with many
natural languages. 2 On the other hand, our knowledge
base comprises three major knowledge levels, consisting
of several independent but interrelated modules:
Lexical level:
• The Lexicon stores morphosyntactic, pragmatic and
collocational information about lexical units.
• The Morphicon helps our system to handle cases of
inflectional morphology.
Grammatical level:
• The Grammaticon stores the constructional
schemata which help Role and Reference Grammar
(RRG) to construct the semantics-to-syntax linking
algorithm (Van Valin and LaPolla, 1997; Van Valin,
2005).
1
We use the name “FunGramKB Suite” to refer to our
knowledge-engineering tool (www.fungramkb.com) and
“FunGramKB” to the resulting knowledge base. FunGramKB
Suite was developed in C# using the ASP.NET 2.0 platform and
a MySQL database.
2
English and Spanish are fully supported in the current version
of FunGramKB Suite, although we have just begun to work with
other languages, such as German, French, Italian, Bulgarian and
Catalan.
Conceptual level:
• The Ontology is presented as a hierarchical
catalogue of the concepts that a person has in mind,
so here is where semantic knowledge is stored in the
form of meaning postulates. The Ontology consists
of a general-purpose module (i.e. Core Ontology)
and several domain-specific terminological modules
(i.e. Satellite Ontologies).
• The Cognicon stores procedural knowledge by
means of scripts, i.e. conceptual schemata in which a
sequence of stereotypical actions is organised on the
basis of temporal continuity, and more particularly
on the basis of Allen's temporal model (1983).
• The Onomasticon stores information about instances
of entities and events, such as Bill Gates or 9/11.
This module stores two different types of schemata
(i.e. snapshots and stories), since instances can be
portrayed synchronically or diachronically.
In the FunGramKB architecture, every lexical or
grammatical module is language-dependent, whereas
every conceptual module is shared by all languages. In
other words, computational linguists must develop one
Lexicon, one Morphicon and one Grammaticon for
English, one Lexicon, one Morphicon and one
Grammaticon for Spanish and so on, but knowledge
engineers build just one Ontology, one Cognicon and one
Onomasticon to process any language input conceptually.
In this scenario, FunGramKB adopts a conceptualist
approach, since the Ontology becomes the pivotal module
for the whole architecture.
The rest of this paper is organized as follows. In sections 2
and 3, we explore most of the modules in the FunGramKB
conceptual, lexical and grammatical levels. In section 4,
we present some of the tools available in FunGramKB
Suite. In section 5, we explore the role of FunGramKB
when it is integrated into an NLP system. Finally, some
conclusions are presented in section 6.
2667
2.
The FunGramKB Conceptual Level
The model of “scheme” originated in cognitive
psychology, and subsequently implemented in artificial
intelligence, is fundamental to the representation of the
world knowledge in FunGramKB. In our knowledge base,
conceptual schemata are classified according to two
parameters: prototypicality and temporality. On the one
hand, conceptual representations can store prototypical
knowledge (i.e. proto-structures) or can serve to describe
instances of entities or events (i.e. bio-structures). For
example, the description of the meaning of song involves
the construction of the proto-structure of its
corresponding concept; however, if we want to provide
information about the song Heartbreak Hotel, then we
should do it through a bio-structure. On the other hand,
knowledge within conceptual schemata can be presented
atemporally (i.e. microstructures) or in a temporal
framework (i.e. macrostructures). For example, the
biography of Elvis Presley requires a macrostructure;
however, a microstructure is sufficient to describe the
profession of singer. Therefore, and as shown in Table 1,
the convergence of the values of these two parameters
results in a typology of four different conceptual schemata
which shape the FunGramKB conceptual level.
TEMPORALITY
P
R
O
T
O
T
Y
P
I
C
A
L
I
T
Y
+
-
+
Proto-microstructure
(Meaning postulate)
Proto-macrostructure
(Script)
(1b) A person or animal moves something towards
themselves with their hand or mouth, providing that
they hold it firmly.
Since the FunGramKB conceptual modules use COREL
as the “common language” for schemata representation,
natural language understanding systems will only require
one “common reasoner”. Indeed, we are currently
developing an automated cognizer with human-like
defeasible reasoning powers which will be able to draw
conclusions from information about facts of the real world
and knowledge from the repository of FunGramKB
meaning postulates, scripts, snapshots and stories. This
reasoner is being implemented by using Drools 5.0
platform, which is provided with a forward-chaining
inference rules engine whose "native" rule language is
powerfully enough so as to preserve the semantic
expressivity of COREL. Moreover, Drools supports
reasoning over temporal relations between events within
an interval-based framework, especially useful for the
FunGramKB macrostructures (cf. section 2.2).4
2.1 The Core Ontology
The FunGramKB Core Ontology is deemed as an IS-A
conceptual hierarchy which allows non-monotonic
multiple inheritance. This ontology is both universal and
linguistically-motivated.
-
Bio-microstructure
(Snapshot)
Bio-macrostructure
(Story)
Table 1: Typology of conceptual schemata in
FunGramKB.
Tulving (1985) stated that long-term memory components
do not work in an isolated way but they interact with each
other in order to facilitate information storage and
retrieval. Therefore, a key factor for successful reasoning
in an NLP system is that all these knowledge schemata
must be represented through the same formal language, so
that information sharing can take place effectively among
all conceptual modules. In FunGramKB, this formal
language is COREL (Conceptual Representation
To
illustrate,
(1a)
presents
the
Language).
COREL-formatted meaning postulate of +PULL_00,
whose natural language equivalent is (1b):3
3
(1a) +((e1: +MOVE_00 (x1: +HUMAN_00 ^
+ANIMAL_00)Agent
(x2:
+CORPUSCULAR_00)Theme
(x3)Location
(x4)Origin
(x5)Goal
(f1:
+HAND_00
^
+MOUTH_00)Instrument (f2: (e2: +SEIZE_00
(x1)Theme (x2)Referent))Condition)(e3: +BE_00
(x1)Theme (x5)Referent))
Periñán-Pascual and Arcas-Túnez (2004) described the formal
grammar of well-formed predications in FunGramKB meaning
Firstly, the Core Ontology takes the form of a universal
concept taxonomy, where “universal” means that every
concept we imagine has, or can have, an appropriate place
in the ontology (Corcho, Fernández López and Gómez
Pérez, 2001). A universal approach is adopted on the
relation between language and conceptualization, where
cross-lingual differences in syntactic constructions do not
necessarily involve conceptual differences (cf. Jackendoff,
1990).
Secondly, the Core Ontology is linguistically motivated,
but not language-dependent. In other words, the Ontology
is involved with the semantics of lexical units, but the
knowledge stored in the Ontology is not specific to any
particular language. In this respect, it is commonly said
that the model of the world portrayed in a particular
ontology is generally biased by distinctions made in the
knowledge engineers’ languages (Hovy and Nirenburg,
postulates.
In fact, Drools 5.0 implements all the temporal operators
defined by Allen’s theory (1983).
4
2668
1992). Consequently, an ontology could be closer to some
language communities than to others, finally affecting the
ontology design. However, this is not a real problem in
FunGramKB, because the structuring of the Ontology is
guided by a process of negotiation.
The FunGramKB Core Ontology distinguishes three
different conceptual levels, each one of them with
concepts of a different type:
(i)
Metaconcepts, e.g. #ABSTRACT, #COLLECTION,
#EMOTION, #POSSESSION, #TEMPORAL etc,
constitute the upper level in the taxonomy. The
analysis of the upper level in the main linguistic
ontologies—DOLCE (Gangemi et al., 2002),
Generalized Upper Model (Bateman, Henschel and
Rinaldi, 1995), Mikrokosmos (Mahesh and
Nirenburg, 1995), SIMPLE (Lenci et al., 2000),
SUMO (Niles and Pease, 2001)—led to a
metaconceptual model whose design contributes to
the integration and exchange of information with
other ontologies, providing thus standardization and
uniformity. The result amounts to forty-two
metaconcepts distributed in three subontologies:
#ENTITY, #EVENT and #QUALITY.
(ii) Basic concepts, e.g. +BOOK_00, +DIRTY_00,
+FORGET_00, +HAND_00, +MOVE_00 etc, are
used in FunGramKB as defining units which enable
the construction of meaning postulates for basic
concepts and terminals, as well as taking part as
selectional preferences in thematic frames. Instead
of adopting a strong approach like that represented
by the Natural Semantic Metalanguage (cf. Goddard
and Wierzbicka, 2002), which identifies a reduced
inventory of semantic primitives that are used to
represent meaning, FunGramKB posits an inventory
of basic concepts which can be used to define any
word in any of the European languages that are
claimed to be part of the Ontology. The starting point
for the identification of our basic concepts was the
defining vocabulary in Longman Dictionary of
Contemporary English (Procter, 1978), though deep
revision was required in order to perform the
cognitive mapping into a single inventory of about
1,300 basic concepts.5
2.2 The Cognicon
The FunGramKB script is structured into one or more
predications within a linear temporal framework—more
particularly, Allen’s interval temporal model (1983). This
model is based on a representation of time as a
partially-ordered graph where nodes represent events and
arcs are tagged with one or more relations of temporal
ordering. In FunGramKB, every predication included in a
script represents an event E which is treated as an interval
consisting of a pair of time points (i, t), i.e. the start
time-point (i) and the end time-point (t). For example,
supposing that an event occurs in the interval E1(i1, t1)
and another event occurs in the interval E2(i2, t2), the
interval relation Before(E1, E2), i.e. the event described
by the predication e1 occurs before the event described by
the predication e2, is subject to the constraint t1 < i2.
Moreover, Allen devised an interval-based constraint
propagation algorithm which computes all temporal
relations taking place when a new event is added to the
graph. For instance, and following the previous example,
if event E3 is added, and E3 occurs during E2, then the
system automatically infers that E1 is before E3. Table 2
presents those interval relations from Allen’s model
which have been incorporated into the Cognicon.
1.
2.
3.
4.
5.
6.
7.
Interval relations
Before(E1, E2)
Meets(E1, E2)
Overlaps(E1, E2)
Starts(E1, E2)
During(E1, E2)
Finishes(E1, E2)
Equals(E1, E2)
Constraints
(t1 < i2)
(i2 = t1)
(i1 < i2) & (i2 < t1) & (t1 < t2)
(i1 = i2) & (t1 < t2)
(i2 < i1) & (t1 < t2)
(i2 < i1) & (t1 = t2)
(i1 = i2) & (t1 = t2)
Table 2: Interval relations in the Cognicon.
To illustrate, (2) presents the first nine predications in the
classical script @EATING_AT_RESTAURANTS, whose
resulting graph is represented in Figure 1.
(2) *(e1: +ENTER_00 (x1: +CUSTOMER_00)Agent
(x1)Theme
(x2)Location
(x3)Origin
(x4:
+RESTAURANT_00)Goal (f1: (e2: +BE_01
(x1)Theme (x5: +HUNGRY_00)Attribute))Reason)
*(e3:
$ACCOMPANY_00
(x6:
+WAITER_00)Agent (x6)Theme (x7)Location
(x8)Origin (x9: +TABLE_00)Goal)
(iii) Terminals, e.g. $AUCTION_00, $VARNISH_00,
$CADAVEROUS_00,
$SKYSCRAPER_00,
$METEORITE_00 etc, are those concepts which
lack definitory potential to take part in the
FunGramKB meaning postulates. The hierarchical
structuring of the terminal level is very shallow.
*(e4: +SIT_00 (x1)Theme (x9)Location)
*(e5: +TAKE_01 (x6)Agent (x10: $MENU_00 |
$WINE_LIST_00)Theme
(x11)Location
(x12)Origin (x9)Goal)
*(e6: +REQUEST_01 (x1)Theme (x13: +FOOD_00
| +BEVERAGE_00)Referent (x6)Goal)
5
This basic level has been tested for validation with the defining
vocabulary in the dictionaries of other languages, e.g.
Diccionario para la Enseñanza de la Lengua Española
(VOX-Universidad de Alcalá de Henares, 1995).
+(e7: +SAY_00 (x6)Theme (x14: (e8: +COOK_00
2669
(x15:
$COOK_D_00)Theme
+FOOD_00)Referent))Referent (x15)Goal)
(x16:
*(e9:
+TAKE_01
(x6)Agent
+BEVERAGE_00)Theme (x18)Location
$BAR_00)Origin (x9)Goal)
(x17:
(x19:
1^2
E1
E3
2
E4
1
E5
1
E6
5
E2
1
E7
1
E8
1
E9
Figure 1: Temporal-knowledge representation in the
FunGramKB scripts.
In the FunGramKB scripts, the nodes in the propositional
networks can represent either predications or script
activators, where the latter include a script identifier and a
list of participant-based mappings from the host script to
the guest script, as shown in (3).
(3) *(e22: @PAY_CASH_00 (x1: x2, x2: f1))
For instance, within the scenario described by
@GOING_SHOPPING_00, we can call another script
describing the method of payment, which turns out to be
also included in the scripts @GOING_TO_
RESTAURANTS_00, @GOING_TO_CINEMAS_00,
etc. In this case, @GOING_SHOPPING_00, @GOING_
TO_RESTAURANTS_00, @GOING_TO_CINEMAS_
00 etc become the host scripts for @PAY_CASH_00,
whose role is that of a guest script.
Concerning the full integration of host and guest scripts,
and following example (3), participants x1 and x2 in the
host script are mapped into x2 and f1 in the guest script
respectively. Thus, the FunGramKB script activators
explicitly state those participants whose referents in the
real world typically coincide.6 Furthermore, it is possible
to call more than one script within the same activator,
provided that the disjunctive logical operator is used, as
can be seen in (4).
(4) *(e22: @PAY_CASH_00 (x1: x2, x2: f1) ^
@PAY_CARD_00 (x1: x1, x2: x2))
In contrast with the semantic knowledge repository of the
6
Although FunGramKB scripts are clearly interconnected
through activators, these macrostructures are not hierarchically
organized as in Schank’s dynamic memory (1982).
Ontology, the possibility of calling a whole script within
another script gives us the chance to introduce
culturally-biased knowledge in the Cognicon, since every
script is assigned a geographical feature determining the
continent, country, etc where that knowledge is typically
true. 7 Unlike Schank and Abelson’s expectation-based
model (1977), which deeply influenced the theoretical
foundation of the Cognicon, FunGramKB is ready to
manage “cultural distinctiveness”, which is commonly
found in procedural knowledge, e.g. social protocols.
2.3 The Onomasticon
The Onomasticon stores information about named entities
and events, i.e. instances of concepts, in the form of
bio-structures. To illustrate, we present some of the
predications in the snapshot (5a) and the story (6a)
assigned to %TAH_MAHAL_00, whose natural language
equivalents are presented in (5b) and (6b) respectively:
(5a) +(e1: +BE_02 (x1: %TAH_MAHAL_00)Theme (x2:
%INDIA_00)Location)
*(e2: +BE_01 (x1)Theme (x3: +WHITE_00 &
$MARBLE_00)Attribute)
*(e3: +COMPRISE_00 (x1)Theme (x4: 1
$DOME_00 & 4 +TOWER_00)Referent)
(6a) +(e1:
past
+BUILD_00
(x1)Theme
(x2:
%TAH_MAHAL_00)Referent (f1: 1633)Time)
+(e2:
past
+BE_00
(x2)Theme
(x3:
%WORLD_HERITAGE_SITE)Referent
(f2:
1983)Time)
(5b) The Tah Mahal is located in India.
Its main material is white marble.
The Tah Mahal has a main dome and four towers.
(6b) The Tah Mahal was built in 1633.
The Tah Mahal became a UNESCO World Heritage
site in 1983.
Unlike other FunGramKB modules, the population of the
Onomasticon is taking place semi-automatically, by
exploiting the DBpedia knowledge base (Bizer et al.,
2009). The DBpedia project 8 is intended to extract
structured information from Wikipedia, turn this
information into a rich knowledge base, which currently
describes more than 2.6 million entities, and make this
knowledge base accessible on the Web. The population
process of the Onomasticon is being performed as
follows:
(i)
7
We are manually creating template-based rules
which can map the knowledge stored in the DBpedia
ontology9 into COREL-formatted schemata.
The “default” value of this feature states that a given script is
universally applicable.
8
http://dbpedia.org
9
The DBpedia ontology, which was manually created from the
most commonly-used Wikipedia infoboxes, takes the form of a
2670
(ii) These rules are deployed to FunGramKB Suite,
where the mapping occurs automatically.
(iii) Since DBpedia automatically evolves as Wikipedia
changes, the Onomasticon will be periodically
updated via web service.
3.
The FunGramKB Lexical and
Grammatical Levels
The FunGramKB lexical model is basically derived from
OLIF 10 (McCormick, 2002; McCormick, Lieske and
Culum, 2004) and enhanced with EAGLES/ISLE
recommendations 11 (Calzolari, Lenci and Zampolli,
2001a, 2001b, 2003; Monachini et al., 2003) with the
purpose of designing robust computational lexica. The
FunGramKB lexical entries, which can be saved as
XML-formatted feature-value data structures, allow the
following types of information:12
-
-
Basic: headword, index, and language.
Morphosyntax: graphical variant, abbreviation,
phrase constituents, category, number, gender,
countability, degree, adjectival position, verb
paradigm and constraints, and pronominalization.
Core Grammar: Aktionsart, lexical template and
construction.
Miscellaneous: dialect, style, domain, example and
translation.
Unlike many other NLP lexical databases, the
FunGramKB lexical and grammatical levels are grounded
in sound linguistic theories, allowing the system to
capture syntactic-semantic generalizations which are able
to provide both explanations and predictions of language
phenomena. Evidently, it is really much easier to build
NLP systems when linguistic theories are neglected, but
NLP applications which can work perfectly with no
foundation in any linguistic theory are deceptively
intelligent (Halvorsen, 1988), since they don’t allow
natural language understanding. In this respect, the
linguistic foundation of FunGramKB is inspired on RRG
and the Lexical Constructional Model (LCM).
On the one hand, RRG is one of
functional models on the linguistic
grammatical
model
communication-and-cognition view
the most relevant
scene today. This
adopts
a
of language, i.e.
shallow IS-A hierarchy of 170 classes, containing 720
properties.
10
OLIF (Open Lexicon Interchange Format) is an
XML-compliant standard for lexical/terminological data
encoding.
11
ISLE (International Standards for Language Engineering),
which is an extension of EAGLES work, supports R&D on
human-language technology issues.
12
Mairal Usón and Periñán-Pascual (2009) presented the
anatomy of the FunGramKB Lexicon by describing the different
types of features which form part of a predicate’s lexical entry.
morphosyntactic structures and grammatical rules should
be explained in relation to their semantic and
communicative functions. In RRG, the semantic and the
syntactic components are directly mapped in terms of a
linking algorithm, which includes a set of rules that
account for the syntax-semantics interface. As a result,
RRG allows an input text to be represented in terms of a
logical structure. For example, the logical structure of the
lexical unit ask for is (7).
(7) [do’ (x, [say’ (x, y)])] PURP [do’ (y, 0)] CAUSE
[BECOME have’ (x, z)]
In FunGramKB, the RRG logical structure has been
enhanced by a new formalism called “conceptual logical
structure” (Periñán-Pascual and Mairal Usón, 2009), so a
logical structure such as (7) is now replaced by the
representation (8).
(8) [do (xTheme, [+REQUEST_01 (xTheme, yGoal)])] PURP
[do (yGoal, 0)] CAUSE [BECOME +REQUEST_01
(xTheme, zReferent)]
The main benefits of CLSs can be summarized as follows:
(i)
CLSs are real language-independent representations,
since they are made of concepts and not words. One
of the consequences of this interlingual approach is
that redundancy is minimized while informativeness
is maximized.
(ii) The inferential power of the reasoning engine is
more robust if predictions are based on cognitive
expectations. In order to perform some reasoning
with the input, the CLS should be transduced into a
COREL representation, so that it can be enriched by
the knowledge in meaning postulates, scripts,
snapshots and stories.
Therefore, CLSs serve to build a bridge between the
FunGramKB conceptual level and the particular
idiosyncrasies coded in a given linguistic expression. For
instance, the sentence Betty asked Bill for an apple has the
CLS (9), which can be mapped into the COREL
representation (10).
(9) <IF DECL <TNS PAST <[do (%BETTY_00Theme,
[+REQUEST_01
(%BETTY_00Theme,
%BILL_00Goal)])] PURP [do (%BILL_00Goal, 0)]
CAUSE
[BECOME
+REQUEST_01
(%BETTY_00Theme, +APPLE_00Referent)]>>>
(10) +(e1:
past
+REQUEST_01
(x1:
%BETTY_00)Theme (x2: +APPLE_00)Referent
(x3: %BILL_00)Goal)
In this CLS-COREL mapping process, the grammatical
operators, the FunGramKB concepts and their thematic
roles are the only CLS elements taken into account.
2671
On the other hand, the LCM (Ruiz de Mendoza and
Mairal, 2008; Mairal and Ruiz de Mendoza, 2009), which
is grounded in the RRG framework, goes beyond the core
grammar. The LCM incorporates meaning dimensions
that have a long tradition in pragmatics and discourse
analysis. Thus, the LCM recognizes the following four
levels of constructional meaning:
(i)
Level 1, or argumental layer, accounts for the core
grammatical properties of lexical items.
(ii) Level 2, or implicational layer, is concerned with the
inferred meaning related to low-level situational
cognitive models (or specific scenarios), which give
rise to meaning implications of the kind that has
been traditionally handled as part of pragmatics
through implicature theory.
(iii) Level 3, or illocutionary layer, deals with traditional
illocutionary force, which is considered a matter of
high-level situational models (or generic scenarios).
(iv) Level 4, or discourse layer, addresses the discourse
aspects, with particular emphasis on cohesion and
coherence phenomena.
In the Grammaticon, each one of these constructional
levels is computationally implemented into a
Constructicon. Thus, the CLS (9) is automatically
generated by means of the Core Grammar of the verb
together with the grammatical information in the
L1-Constructicon. Furthermore, CLSs can be
incrementally expanded by each type of Constructicon.
For instance, an L3-CLS is that logical structure which
has been enriched by the implicational and illocutionary
levels of constructional meaning.
Currently most NLP lexical databases—e.g. SIMPLE
(Lenci et al., 2000) or EuroWordNet (Vossen, 1998),
among many others—adopt a relational approach to
represent lexical meanings, since it is easier to state
associations among lexical units in the way of meaning
relations than describing the conceptual content of lexical
units formally. However, although large-scale
development of deep-semantic resources requires a lot of
time, effort and expertise, not only is the expressive power
of conceptual meanings much more robust, but the
management of their knowledge also becomes more
efficient (cf. Periñán-Pascual and Arcas-Túnez, 2007).
4.
Tools in FunGramKB Suite
FunGramKB Suite is provided with a set of user-friendly
tools to browse, check and edit the knowledge base. To
illustrate, some of these tools are briefly described:
(i)
Conceptual, lexical and grammatical modules can be
browsed via a GUI, displaying specific feature-value
information about their elements.
(ii) When building conceptual knowledge in the form of
meaning postulates, scripts, snapshots or stories, a
syntactic-semantic validator is triggered, so that
consistent well-formed constructs can be stored.
(iii) In order to help knowledge engineers to determine
the granularity of meaning postulates, a checklist
suggests the semantic components which could
become relevant on the basis of the conceptual
dimension to which the concept belongs.
5.
Integrating FunGramKB into an NLP
System
One of the first attempts to integrate FunGramKB into an
NLP system is aimed at improving the performance of
UniArab, an Arabic-to-English machine translator (Nolan
and Salem, 2009; Salem, Hensman and Nolan, 2008a,
2008b; Salem and Nolan, 2009a, 2009b). The advantage
of UniArab lies in the deployment of an interlingua
architecture which uses a robust functional linguistic
model founded on RRG in the machine translation kernel.
On the one hand, UniArab is built upon an interlingua
machine translation architecture, which is more flexible
and scalable for multilingual generation. On the other
hand, one of the primary strengths of UniArab is the
accurate representation of the RRG logical structure of an
Arabic sentence. To illustrate, the logical structure (11) is
built from the Arabic sentence (12), whose translation
into English is sentence (13).
(11) <TNS:PAST[do'(Khalid,[read'(Khalid,( book)])]>
(12)
(13) Khalid read the book.
Currently, UniArab covers a representative broad
selection of words and can translate simple sentences
including intransitive, transitive and ditransitive clauses,
as well as copular-like nominative clauses. Concerning
the evaluation of UniArab, this system clearly
outperforms existing machine translators in the
processing of simple sentences, suggesting that RRG is a
promising candidate for interlingua-based machine
translation. In fact, Salem and Nolan (2009b)
demonstrated that UniArab provides more accurate and
grammatically-correct translations than statistical
machine translators such as Google (2009) and Microsoft
(2009).
However, the model of UniArab devised by Brian Nolan
and his research team fails to provide an adequate
treatment of the semantics of lexical units. For instance,
UniArab avoids the problem of word sense
disambiguation by adopting a naive one-word-one-sense
approach to lexical polysemy. To overcome this problem,
among many others, the UniArab lexical database is
replaced by FunGramKB, where lexical entries are more
2672
informative and meaning capabilities are deeper. Thus, at
the end of the syntax-semantics processing, the enhanced
version of UniArab generates a syntactic representation of
the input where lemmas have been replaced by conceptual
tags. In the case of sentence (12), the output would be the
parenthetical representation (14), whose concepts are also
provided with lexico-conceptual information represented
as feature-value matrices.
(14) S(NP(n(%KHALID_00)),
VP(v(+READ_00),
NP(det(the), n(+BOOK_00))))
The RRG logical structure (11) is then developed out of
the phrasal structure (14), but now taking the form of the
CLS (15).
(15) <IF DECL <TNS PAST < do ($KHALID_00Theme,
[+READ_00
($KHALID_00Theme,
+BOOK_00Referent)]
&
INGR
+READ_00
(+BOOK_00Referent)>>>
The shift from the standard RRG model of logical
structure to the CLS approach opens a new avenue for
UniArab to cope with complex multilingual input.
6.
Conclusions
This paper addresses the three levels of knowledge which
shape FunGramKB, a multipurpose knowledge base for
NLP systems. We highlight two main contributions of our
project in comparison with other similar knowledge bases.
On the one hand, the FunGramKB conceptual level
enables the full integration of semantic, procedural and
episodic knowledge by sharing both the knowledge
representation language and the reasoning engine. As a
result, expectations on the occurrence of typical events in
a given situation are based on COREL schemata, a
concept-oriented interlingua whose inferential power is
greater than the traditional approach to lexical semantics.
On the other hand, the FunGramKB lexico-grammatical
levels are grounded in a solid linguistic theory in order to
capture syntactic-semantic generalizations which can
manage and interpret data. In this respect, both the RRG
and the LCM frameworks inspired the construction of the
CLS, a lexically-driven interlingua through which the
system is able to predict a wide range of linguistic
phenomena (e.g. passivization) in the language generation
process. Whereas the CLS serves as the pivot language
between the input text and the COREL representation, the
latter serves as the pivot language between the CLS and
the automated reasoner. Consequently, the primary goal of
our project is the development of an NLP knowledge base
sufficiently robust to help language engineers to design
intelligent natural language understanding systems.
7.
Acknowledgements
Financial support for this research has been provided by
the DGI, Spanish Ministry of Education and Science,
grant FFI2008-05035-C02-01/FILO. The research has
been co-financed through FEDER funds.
8.
References
Allen, J. (1983). Maintaining knowledge about temporal
intervals. Communications of the ACM, 26(11), pp.
832--843.
Bateman, JA., Henschel, R., Rinaldi, F. (1995). The
Generalized Upper Model 2.0. Technical report.
IPSI/GMD, Darmstadt.
Bizer, C., Lehmann, J., Kobilarov, G., Auer, S., Becker, C.,
Cyganiak, R., Hellmann, S. (2009). DBpedia:
a crystallization point for the Web of Data. Journal
of Web Semantics: Science, Services and Agents on the
World Wide Web, 7, pp. 154--165.
Calzolari, N., Bertagna, F., Lenci, A., Monachini, M.
(Eds.) (2003). Standards and Best Practice for
Multilingual Computational Lexicons and MILE.
Deliverable D2.2-D3.2. ISLE Computational Lexicon
Working Group.
Calzolari, N., Lenci, A., Zampolli, A. (2001a). The
EAGLES/ISLE computational lexicon working group
for multilingual computational lexicons. In
Proceedings of the First International Workshop on
Multimedia Annotation. Tokyo.
Calzolari, N., Lenci, A., Zampolli, A. (2001b).
International standards for multilingual resource
sharing: the ISLE Computational Lexicon Working
Group. In Proceedings of the ACL 2001 Workshop on
Human Language Technology and Knowledge
Management. Morristown, NJ: Association for
Computational Linguistics, pp. 71--78.
Corcho, O., Fernández López, M., Gómez Pérez, A.
(2001). Technical Roadmap v.1.0. IST-OntoWeb
Project, Universidad Politécnica de Madrid.
Gangemi, A., Guarino, N., Masolo, C., Oltramari, A.,
Schneider, L. (2002). Sweetening ontologies with
DOLCE. In Proceedings of the 13th International
Conference on Knowledge Engineering and
Knowledge Management. Ontologies and the Semantic
Web. London: Springer, pp. 166--181.
Goddard, C., Wierzbicka, A. (Eds.) (2002). Meaning and
Universal Grammar. Amsterdam: John Benjamins.
Google
(2009).
Google
Translator.
http://translate.google.com.
Hovy, E., Nirenburg, S. (1992). Approximating an
interlingua in a principled way. In The DARPA Speech
and Natural Language Workshop, New York.
Jackendoff, R. (1990). Semantic Structures. Cambridge,
Mass.: MIT Press.
Lenci, A., Bel, N., Busa, F., Calzolari, N., Gola, E.,
Monachini, M., Ogonowski, A., Peters, I., Peters, W.,
Ruimy, N., Villegas, M., Zampolli, A. (2000). SIMPLE:
a general framework for the development of
multilingual lexicons. International Journal of
Lexicography, 13(4), pp. 249--263.
Mahesh, K., Nirenburg, S. (1995). Semantic classification
for practical natural language processing. In 6th ASIS
SIG/CR Classification Research Workshop: An
Interdisciplinary Meeting. Chicago: ASIS, pp. 79--94.
Mairal Usón, R., Periñán-Pascual, C. (2009). The
anatomy of the lexicon component within the
2673
framework of a conceptual knowledge base. Revista
Española de Lingüística Aplicada, 22, pp. 217--244.
Mairal Usón, R., Ruiz de Mendoza, F. (2009). Levels of
description and explanation in meaning construction. In
C.S. Butler & J. Martín Arista (Eds.), Deconstructing
Constructions.
Amsterdam/Philadelphia:
John
Benjamins, pp. 153--198.
McCormick, S. (2002). The Structure and Content of the
Body of an OLIF v.2.0/2.1. The OLIF2 Consortium.
McCormick, S., Lieske, C., Culum, A. (2004). OLIF v.2:
A Flexible Language Data Standard. The OLIF2
Consortium.
Microsoft
(2009).
Microsoft
Translator.
htttp://www.windowslivetranslator.com/Default.aspx.
Monachini, M., Bertagna, F., Calzolari, N., Underwood,
N., Navarretta, C. (2003). Towards a Standard for the
Creation of Lexica. ELRA European Language
Resources Association.
Niles, I., Pease, A. (2001). Origins of the Standard Upper
Merged Ontology: a proposal for the IEEE Standard
Upper Ontology. In Working Notes of the IJCAI-2001
Workshop on the IEEE Standard Upper Ontology.
Seattle.
Nolan, B., Salem, Y. (2009). UniArab: an RRG
Arabic-to-English machine translation software. In
Proceedings of the Role and Reference Grammar
International Conference. Berkeley.
Periñán-Pascual, C., Arcas-Túnez, F. (2004). Meaning
postulates in a lexico-conceptual knowledge base. In
Proceedings of the 15th International Workshop on
Databases and Expert Systems Applications. Los
Alamitos, CA: IEEE, pp. 38--42.
Periñán-Pascual, C., Arcas-Túnez, F. (2007). Deep
semantics in an NLP knowledge base. In Proceedings
of the 12th Conference of the Spanish Association for
Artificial Intelligence, Salamanca, pp. 279--288.
Periñán-Pascual, C., Mairal Usón, R. (2009). Bringing
Role and Reference Grammar to natural language
understanding. Procesamiento del Lenguaje Natural,
43, pp. 265--273.
Procter, P. (Ed.) (1978). Longman Dictionary of
Contemporary English. Harlow: Longman.
Ruiz de Mendoza Ibáñez, F., Mairal, R. (2008). Levels of
description and constraining factors in meaning
construction: an introduction to the Lexical
Constructional Model. Folia Linguistica, 42(2), pp.
355--400.
Salem, Y., Hensman, A., Nolan, B. (2008a).
Implementing Arabic-to-English machine translation
using the Role and Reference Grammar linguistic
model. In Proceedings of the 8th Annual International
Conference on Information Technology and
Telecommunication, Galway.
Salem, Y., Hensman, A., Nolan, B. (2008b). Towards
Arabic to English machine translation. ITB Journal, 17,
pp. 20--31.
Salem, Y., Nolan, B. (2009a). Designing an XML lexicon
architecture for Arabic machine translation based on
Role and Reference Grammar. In Proceedings of the
2nd International Conference on Arabic Language
Resources and Tools, Cairo.
Salem, Y., Nolan, B. (2009b). UniArab: a universal
machine translator system for Arabic Based on Role
and Reference Grammar. In Proceedings of the 31st
Annual Meeting of the Linguistics Association of
Germany.
Schank, R.C. (1982). Dynamic Memory. New York:
Cambridge University Press.
Schank, R., Abelson, R.P. (1977). Scripts, Plans, Goals
and Understanding. Hillsdale: Lawrence Erlbaum.
Tulving, E. (1985). How many memory systems are there?
American Psychologist, 40, pp. 385--398.
Van Valin, R.D. Jr. (2005). The Syntax-SemanticsPragmatics Interface: An Introduction to Role and
Reference
Grammar.
Cambridge:
Cambridge
University Press.
Van Valin, R.D. Jr., LaPolla, R. (1997). Syntax, Structure,
Meaning and Function. Cambridge: Cambridge
University Press.
Vossen, P. (1998). Introduction to EuroWordNet.
Computers and the Humanities, 32(2-3), pp. 73--89.
VOX-Universidad de Alcalá de Henares. (1995).
Diccionario para la Enseñanza de la Lengua Española.
Barcelona: Bibliograf.
2674